sparse.txtmedint.sgrlasso.largep_omega: Conduct sparse mediation for large p ( p > n)

In seonjoo/smm: Sparse Mediation For High-Dimensional Mediators

Description

Sparse mediation with sparse group lasso (for mediation paths) and sparse precision matrix estimation using fast computation of inverse matrix

Usage

sparse.txtmedint.sgrlasso.largep_omega(
  X,
  M,
  Y,
  X.scale = TRUE,
  tol = 10^(-10),
  max.iter = 10,
  lambda1 = exp(seq(1, -3, length = 10)),
  lambda2 = exp(-1),
  alpha = c(0.95, 0.5),
  group.penalty.factor = c(1, rep(1, ncol(M))),
  penalty.factor = c(1, rep(1, ncol(M) * 3)),
  verbose = FALSE,
  Omega.out = FALSE,
  threshold = 0,
  non.zeros.stop = ncol(M)
)

Arguments

X	One-dimensional predictor
M	Multivariate mediator
Y	Outcome
tol	(default -10^(-10)) convergence criterion
max.iter	(default=100) maximum iteration
lambda1	(default=seq(0.02,0.4,length=5)) tuning parameter for regression coefficient L1 penalization
lambda2	Tuning parameter for Covariance matrix L1 penalization
alpha	(defult=1) tuning parameter for L2 penalization
group.penalty.factor	(V+1)-dimensional group penalty factor vector. If a user does not want to penalize mediator, specify 0 otherwise 1. The first element is the direct effect followed by V-mediators. The default value is c(0,rep(1,V)).
penalty.factor	(1+3*V)-dimensional sparsity penalty factor vector. The order of parameters: c(c,b11,b12,...,b1V, b21,...,b2V, a1,a2,...,aV)
verbose	(default=FALSE) print progress
Omega.out	(defult=FALSE) output Omega estimates (beta version WIP.)
threshold	(default=10^(-5))
non.zeros.stop	(default=ncol(M)) When to stop the regularization path.

Details

Fit a mediation model via penalized maximum likelihood and structural equation model. The regularization path is computed for the lasso or elasticnet penalty at a grid of values for the regularization parameter lambda. Currently, mediation analysis is developed based on gaussian assumption.

Multiple Mediaton Model: (1) M = Xa + e1 (2) Y = Xc' + Mb + e2 And in the optimization, we do not regularize c', due to the assumption of partial mediation.

Value

c directeffect

hatb Path b (M->Y given X) estimates

hata Path a (X->M) estimates

medest Mediation estimates (a*b)

alpha

lambda1 Tuning parameters for regression coefficients

lambda2 Tuning parameters for inversed covariance matrix (Omega)

nump Number of selected mediation paths

Omega Estimated covariance matrix of the mediator

Author(s)

Seonjoo Lee, [e-mail]

References

TBA

Examples

library(sparsemediation)
N=200
V=100
set.seed(1234)
a = c(rep(0,V-3),rep(1,3))*5;
b1=b2=rep(0,V)
b1[V]<-5
b2[V]<-5
X = rnorm(N)
M =  X %*% t(a)+ matrix(rnorm(N*V),N,V)
Y =  X + M %*% b1 + (X*M) %*% b2 +  rnorm(N)
#fit=sparse.txtmedint.sgrlasso.largep_omega(X,M,Y)

seonjoo/smm documentation built on Feb. 11, 2021, 5:54 a.m.